IEX is currently one of about 40 active ATSs, most of which would be considered dark pools. Collectively, ATSs trade roughly 20% of overall U.S. equity volume. Most of these dark pools operate as continuous order books, much like the lit exchanges (NYSE, Nasdaq, etc.), and most of them are operated by brokers. From a trading perspective, the major differences between continuous dark pools and exchanges are as follows:
- Dark pools do not disseminate quotes; all orders in dark pools are fully hidden. Most if not all exchanges allow hidden orders as well, but these represent a small proportion of matched volume (e.g. hidden orders represent ~5% of volume on Arca). Note that as of April 1, IEX has started disseminating quotes for orders designated to be displayed, but for the purposes of this post, we will only be discussing hidden order types on IEX, which as of this writing represent the vast majority of our trading volume.
- Resting orders on dark pools are not protected under Rule 611 of Reg NMS and can thus be traded through, meaning an order can be executed on another market center even if the dark pool order was willing to trade at a better price.
- Dark pools are not required to provide fair access; they may turn away potential customers at their sole discretion. Exchanges, on the other hand, must provide access to any registered broker-dealer that wishes to become a member. Note that most of the major proprietary automated trading firms are registered as broker-dealers.
- Conversely, a dark pool operator can choose to provide access directly to traders and investors; exchanges can only be accessed by registered broker-dealers and cannot be accessed by end investors/traders directly.
Why Trade in a Dark Pool?
There are many reasons why a broker might send an order to a dark pool (keep in mind that most of the major institutional brokers operate their own dark pools - think of all the potential synergies), but I'm going to focus on the reasons why an institutional investor (mutual fund, pension fund, hedge fund, etc.) might want to trade in a dark pool:
- Expose a large order to the market without impacting the price of the stock (hopefully). An institutional investor can throw an order in a dark pool and have a chance at getting a huge execution without pinning the price of the stock at their limit price. Of course, there is no guarantee that the order will execute at all, so there is certainly a trade-off.
- Price improvement / midpoint trading. A decent amount of dark pool volume happens at the midpoint of the NBBO, meaning an investor may get an immediate execution while only having to pay half the spread (about 70% of overall volume on IEX occurs at the mid).
- Counter-party filtering/selection. Some investors like the prospect of choosing with whom they trade. Many dark pool operators limit which firms are allowed to access their pool (e.g. some pools claim to completely disallow aggressive HFTs). Alternatively, some dark pools allow customers to explicitly select which types of counterparties they wish to interact with (e.g. other buyside flow only). Note that IEX operates as a fair access venue, just like the registered exchanges, so we don't offer any of these services, but many ATSs do.
In practice, the second reason (price improvement) is probably the only one that's relevant to smaller firms and retail investors.
The standard order type used by institutional investors in dark pools is midpoint peg, an order type managed by the dark pool operator that simply floats with the market at the midpoint of the NBBO. Institutional brokers and investors face two main challenges when trading with midpoint peg orders in dark pools:
- Orders getting "sniffed out." The whole point of resting a large midpoint peg order in a dark pool is that it won't pin the stock at your limit price. If you're a buyer with a $10.00 top, the price of the stock on the exchanges can freely move around-- it could even drop to say $9.50-- without your order holding the market at $10.00. The dark pool will simply adjust the price of your order up and down (but no higher than $10.00) as the midpoint of the NBBO changes. BUT if anyone is able to detect that your large order is willing to peg all the way up to $10.00, there's a very clear arbitrage opportunity where they buy all the stock available on the exchanges up to $10.00, move the midpoint up to at least $10.00, and then sell to you in the dark pool at $10.00 (and rinse and repeat), and you have failed to accomplish your goal.
- Adverse selection. Adverse selection is the standard industry term to describe the scenario where you execute a trade immediately prior to the market moving to a more favorable price. Adverse selection is analogous to buying a plane ticket on Kayak, and then seeing the exact same flight available for $50 cheaper two days later-- pretty frustrating. In dark pools, this often happens on the scale of single-digit milliseconds or less. The below charts from an ITG paper from 2009 (this is by no means a recent phenomenon) illustrate adverse selection.
In the above chart, the blue line represents the National Best Bid across the exchanges, and the red line represents the National Best Offer. The large midpoint peg buy order floats in the dark pool, executing in pieces in line with the rest of the market throughout the day (presumably smaller-sized sell orders are entering the dark pool at various points, and the large buyer's order has not been fully satisfied until the end of the chart). This is the normal, desirable dark pool experience.
In the second chart, the large buy order only executes in pieces at the local highs immediately before the market moves lower, and as a result its average fill price is substantially worse.
How does this happen? Two main ways:
- Getting "run over" by a larger seller. Suppose a new, even larger, aggressive seller enters the dark pool and trades with the resting buyer, and then still has more shares for sale. If that seller immediately thereafter hits bids on the exchanges, they can push the price of the stock lower, causing an outcome of adverse selection for the dark pool buyer. Getting run over on any market (exchange or dark pool) is one of the main challenges that market makers face, but for a large institutional investor, it's not the worst outcome in the world. The buyer gets their entire order filled in line with the market, so even though they may have been better off had they priced their order slightly lower, it's a fairly positive result overall.
- Stale quote arbitrage. Now suppose instead that a new seller enters the market on the exchanges, pushing the price down, without first checking the dark pool. There is a brief moment in time, after that seller has affected the price on the exchanges, before the dark pool has updated the prices its pegged orders. This window may only last a couple of milliseconds or less, depending on how fast the dark pool processes market data updates, but it creates a brief arbitrage opportunity where a trader can send an exploratory short sell order at the old, higher midpoint price in the dark pool and, if it executes, buy back those shares from the new, lower-priced seller on the exchanges.
Stale Quote Arbitrage
We call this latter case a type of structural arbitrage; it only exists due to a technical limitation of the dark pool (the dark pool can't update its prices fast enough). Whereas a broker can build anti-gaming logic into their trading algorithms to try to prevent orders from getting "sniffed out" (e.g. by piecing the order in smaller pieces or with extra-conservative limit prices), when it comes to structural arbitrage on a dark pool or exchange, there's really not much the broker can do other than simply not trading in that market. And if every single dark pool has a similar exposure to this situation, well then, there's not much recourse.
Ok. So everything up until this point has been theoretical. Prior to IEX, my job was to design and program equity execution algorithms at RBC, a large Canadian bank. The first strategy I ever built was a dark pool aggregator, which would take an institutional investor's order (our client) and split it up among several dark pools, shuffling around seeking out the other side of the trade. The resting midpoint orders we sent to dark pools were certainly adversely selected regularly, and we believed the primary cause to be stale quote arbitrage, but we couldn't definitively confirm this belief-- there was no way for us to synchronize the true state of the market with the dark pool's view at the time of the execution. Still, even though we couldn't be 100% certain of the dynamics at play, we figured our team would be able to solve this and other challenges faced by institutional brokers and investors, and so we left to start IEX.
Preventing Structural Arbitrage
At IEX, one of our top priorities right out of the gate was to prevent structural arbitrage on our market; we did not want our technical limitations to expose any of our customers to sub-optimal trades. To specifically address stale quote arbitrage-- the situation where our system is unable to update the prices of resting pegged orders fast enough to prevent them from getting picked off-- we came up with the idea to impose a tiny delay on all inbound orders. We measured that all-in, it takes us a little over 300 microseconds to recognize a price move on the exchanges and update our pegged orders accordingly, so we introduced a delay on inbound orders of 350 microseconds, so that even if a trader could instantaneously recognize a price change in the market and send us an exploratory order at the old price, their order would be delayed for just long enough to ensure that we would know the new price before their order could execute.
This is one of the most common misconceptions about IEX's "speed bump"-- we're not trying to pick winners and losers or equalize everyone's technology. The trader that gets to our entry ports first will get to trade first, so for the majority of trading strategies, we're simply moving the race from our matching engine to another data center. The purpose of the speed bump is simply to ensure that we don't allow trades at stale prices after the market has already moved. A bookie probably shouldn't allow bets to be placed on a race that's already over, even if they have unknowing customers willing to take the losing side of that bet. Similarly, we impose this delay to ensure that no customer can make a trading decision based on more up-to-date market data than we have ourselves, so that we can better enforce the spirit of Reg NMS.
We went live in late October 2013, and for the first several months there was little-to-no adverse selection on IEX. Of course, there will always be some baseline level of short-term adverse selection from market fluctuations and orders getting run over, but we seemed to have solved the stale quote arbitrage problem.
But then suddenly, and quite dramatically, the incidence of adverse selection on IEX rose in mid-2014:
The above chart shows the % of shares added at the midpoint such that the midpoint was at a better price 10 milliseconds later. From March to July of 2014, the incidence of adverse selection of midpoint orders on IEX rose from about 3-4% to about 9-11%. Given our architecture, stale quote arbitrage should have been impossible, so what was happening?
We realized our speed bump was effective at preventing after-the-fact stale quote arbitrage-- that is, no trader can observe a price move in the market and then effect a trade on IEX at the old midpoint. These new incidences of adverse selection, however, were occurring as much as 1 to 2 milliseconds prior to an NBBO update, far enough in advance that our speed bump wasn't relevant.
This was a bit of an "aha" moment for us. We realized that NBBO changes are not always instantaneous events; rather, they are a coordinated series of events that happen across all of the exchanges in rapid succession. In other words, when a liquid stock moves from $9.99 bid/$10.00 offered to $9.98/$9.99, all of the individual buy orders at $9.99 across all of the exchanges are either filled or canceled, and one by one each exchange's best bid drops, and then the offers fill in. The whole process can take a few milliseconds, but while it's in motion, a trader can see the dominoes falling and have a fairly high degree of confidence that the market is moving.
We briefly considered lengthening our speed bump delay to prevent this new type of pseudo-arbitrage, but we felt the order-of-magnitude increase that would be necessary might be too disruptive to normal trading activity. Then we figured we could introduce a new kind of midpoint peg order that monitors for when the market is in transition and, in those moments, becomes less aggressive, and this is exactly what we did.
The New Order Type: Discretionary Peg
We called the new order type discretionary peg, and it went live in November 2014. Discretionary peg is an order type that is willing to trade as aggressively as the midpoint of the NBBO the vast majority of the time, but in moments when IEX observes the market moving in the order's favor, it is only willing to trade on the passive side of the bid/offer. Just like a fast trader conducting this pseudo-arbitrage, we constantly observe the market for a signal that the price of a stock is in transition, and we use this signal to prevent a trade from occurring at the soon-to-be-stale price.
As an aside, I've heard the question several times going all the way back to our RBC days: if we are able to identify profitable signals like this one, why don't we start a proprietary trading shop for ourselves and just print money? There are many reasons why we started IEX instead of going down this path, but from a practical perspective, it's simply not our competitive advantage. Many arbitrage or pseudo-arbitrage strategies are extraordinarily simple from a trading logic perspective-- this stuff isn't rocket science-- and this case is no different. The challenge is that only one trader can successfully take advantage of each arbitrage opportunity, and to win the race consistently requires extremely fast technology and frequent upgrade. I have no reason to believe our team would be nearly as effective in that space.
Protecting our Resting Orders
So given all this, how could we possibly identify that the market is in transition early enough to help our resting order avoid getting picked off? First, keep in mind that we do still have the speed bump working for us, so we don't need to be the single fastest at picking up the signal-- as long as we can identify that the market is transitioning within 350 microseconds of the very fastest trader, we can protect our resting discretionary peg orders. It turns out that 350 microseconds is an enormous head start, and it makes our job a lot easier. Secondly, the downside of a false positive for IEX is smaller than the downside of a false positive for an arbitrage trader: if a trader has a false positive, and they execute a trade anticipating a market move that doesn't actually happen, they now have a position that they will most likely wind up closing at a loss. If IEX, on the other hand, thinks the market is moving, but it doesn't, we just return the order to its normal behavior. There is a small chance that the discretionary peg order may miss a desirable midpoint trade in this tiny window, so false positives are still not a good thing-- they're just not as bad, so we can afford to be a bit more aggressive with our signal than an arbitrage trader.
Ultimately, IEX doesn't need to win the prediction arms race. Of course we will continue to strive to make our signal as precise as possible, but even if the signal is a little bit crude and noisy, as long as we take away the really obvious profitable scenarios, it should make the entire practice much less desirable to conduct.
Here are the results so far:
Whereas midpoint pegged orders have been seeing adverse selection in the 9-11% range on average since last July, discretionary peg orders are down in that 3-4% range that we saw in the early days of IEX. It's important to note that there is a trade-off between using the two order types: midpoint peg orders earn higher priority than discretionary peg inside the spread, and of course, discretionary peg orders naturally face a lower fill rate by avoiding a subset of trades (1-2% lower in practice so far). All-in-all, however, discretionary peg does seem to be a compelling order type for an institutional broker/investor concerned with short-term adverse selection, and we are very happy with the results to date.
In closing, it is the broker's job to navigate the market effectively on behalf of their customers, but if an exchange or a dark pool has a blind spot that allows for structural arbitrage, there isn't much the broker can do. You can't blame a trader for trying to profit off an inefficiency, but we believe that exchanges and ATSs have the responsibility to ensure that they don't have any blind spots.
Dan is a co-founder and quantitative developer at IEX. He is responsible for building and evolving core functionality for the IEX trading venue, namely its matching engine, smart order router and its newest order type: Discretionary Peg. Forbes recognized Dan as one of their 2015 "30 Under 30".